Skip to content

TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

TwinMarket is a multi-agent LLM simulation framework for stock markets that uses the Belief-Desire-Intention (BDI) cognitive architecture to model heterogeneous investor behavior within a dynamic social network. It demonstrates that LLM agents can reproduce four canonical financial market stylized facts (fat tails, leverage effect, volume-return correlation, volatility clustering) more faithfully than rule-based ABM baselines. Scaling to 1,000 agents yields a monotonic improvement in predictive accuracy against real SSE 50 index data.

Problem

Traditional rule-based Agent-Based Models (ABMs) assume homogeneous agents with static behavioral rules, failing to capture irrational decision factors (overconfidence, loss aversion, herding) and the social interaction dynamics that drive emergent market phenomena like bubbles and crashes. Existing LLM-based financial simulations (CompeteAI, EconAgent, ASFM) either fix agent prompts, ignore inter-agent interactions, or operate at very small scales, obscuring the causal path from individual cognition to systemic emergence.

Method

Each agent is structured around a BDI loop: Beliefs are agent-specific system prompts encoding demographic and behavioral bias profiles synthesized from 639 real Xueqiu user profiles and CSMAR transaction data; beliefs are updated daily via a sentiment score (1–5) across five self-assessed dimensions. Desires drive autonomous queries to retrieve relevant stock/news information. Intentions produce final trading orders and social media posts/reposts.

The macro environment consists of (1) an order-driven trading system over SSE 50 stocks and (2) a social media platform. The social network is a dynamic graph G=(V,E) where edge weights are weighted Jaccard similarities over time-decayed trading intensities (exponential decay λ). Information propagation uses a hot-score ranking: h(p) = log10(upvotes−downvotes+1) / (T_p+1), where T_p decays with ~2-week half-life, and each agent receives the top-k posts from similarity-threshold neighbors per day. Simulations run from June 2023 with GPT-4o powering each agent.

Key Contributions

  • BDI-structured LLM agents with calibrated heterogeneous behavioral biases grounded in real Chinese retail investor data
  • Dynamic social network with time-decayed edge weights and a personalized hot-score information aggregation mechanism
  • Empirical replication of four financial stylized facts, outperforming HPM and BH rule-based ABM baselines on all metrics
  • Emergence analysis demonstrating self-fulfilling belief-price cycles, opinion-leader polarization, and rumor-driven market turbulence
  • Demonstrated scaling law: monotonically decreasing RMSE/MAE as the fraction of actively trading agents increases (10%→80%), validated at 1,000 agents

Results

  • Stylized facts vs. real data (kurtosis / neg-return autocorrelation / GARCH α+β): Real=7.26/0.14/0.95; TwinMarket=5.24/0.11/0.89; ABM-HPM=4.47/0.05/0.82; ABM-BH=4.99/0.19/0.72 — TwinMarket closest to real on kurtosis and GARCH persistence
  • Ablation: Full model RMSE=0.02, MAE=0.02, Corr=0.77; w/o BDI: RMSE=0.07, Corr=0.34; w/o Heterogeneity: RMSE=0.09, Corr=−0.61
  • Wealth inequality: Gini coefficient rises over the 5-month simulation; top 10% agents average +6.65% return at 4.02% turnover; bottom 50% average −10.52% at 7.03% turnover
  • Rumor experiment: Sell/buy ratio rises from 0.495 to 0.997 under rumor exposure; rumor-exposed agents assign asset valuations 27.5% lower than baseline
  • Scaling: RMSE and MAE decrease monotonically from 10% to 80% agent activation rate

Limitations

  • Scoped exclusively to China A-share market (SSE 50 index); generalizability to other market structures unvalidated
  • Simplified single daily call-auction mechanics rather than continuous trading
  • Zero-sum environment assumption
  • Validation uses 100 agents for stylized-fact experiments; scalability to larger populations shown but not fully validated against the same statistical benchmarks
  • Relies on GPT-4o, introducing API cost and reproducibility concerns at scale
  • Social network construction assumes trading-similarity as proxy for social connection, which may not reflect real information-sharing patterns

Relevance to Agentic AI / LLM Agents

TwinMarket is a high-fidelity instantiation of the BDI agentic architecture at scale, demonstrating that cognitive structure (belief updating, goal-directed querying, intention-driven action) is empirically necessary for emergent realism — removing BDI collapses correlation from 0.77 to 0.34. The framework directly addresses open questions in multi-agent LLM research: how to model persistent agent memory, heterogeneous persona injection, information filtering in networked environments, and the micro-to-macro emergence that arises from thousands of interacting agents. Its ablation results provide rare quantitative evidence for why structured cognition and diversity matter in LLM agent design, relevant to anyone building realistic multi-agent simulations for social, economic, or organizational domains.